Warn that enu_to_body/body_to_enu take origins in different frames

[qmohsu]

Summary

core/coords/transforms.py's enu_to_body (line 454) and body_to_enu
(line 485) are a true inverse pair only when the two origin arguments are
converted into each other, not when the same array is passed to both:

• enu_to_body(..., body_origin_enu=...) wants the body origin expressed
  in ENU.
• body_to_enu(..., enu_origin_body=...) wants the ENU origin expressed
  in the body frame.

They're related by enu_origin_body = -C @ body_origin_enu (C = the
rotation matrix for the same roll/pitch/yaw), not equality. Passing one
array to both — the obvious thing a reader would try — silently returns a
wrong point.

Measured before the change, with x=[10,20,5], origin=[1,2,3],
roll,pitch,yaw=(0.1,0.2,0.3), same array to both origin arguments:

• Full 3D round-trip error: 7.4727 m
• Up-axis component alone: 6.2210 m (this is what the task brief's
  "6.22 m" referred to — a per-axis figure, not the vector norm; both
  numbers are reported here since they differ).
• With the correct conversion instead: round-trip error is 3.55e-15 m
  (float noise).

Nothing in the repo previously exercised this: TestEnuBody::test_round_trip
only ever called with the default origin=None, and no caller in the repo
passes a non-default origin to either function.

Scope decision

Signatures are unchanged — they mirror the book's Eq. (2.6)/(2.7), and this
repo's convention is code follows the book unless the book is wrong (see
docs/book_errata.md). This is a documentation + test fix only:

1. core/coords/transforms.py:454 and :485 — added a Notes: section
   with a CONVENTION WARNING to both docstrings, naming the frame of each
   origin argument and giving the conversion. Style matches
   core/sensors/strapdown.py's quat_to_rotmat CONVENTION WARNING (the
   existing house style for this kind of cross-function frame mismatch note).
2. tests/core/coords/test_transforms.py — two new tests in TestEnuBody:
   • test_round_trip_with_offset_origin: round-trips through a non-None
     origin using the correct -C @ conversion.
   • test_same_array_as_both_origins_does_not_round_trip: an inequality
     assertion (with an explanatory comment) pinning that passing the same
     array to both is a trap — if a future change unifies the two arguments,
     this goes green and should read as the convention having changed.
3. Checked docs/ch2_equation_mapping.md and docs/CH2_QUICK_REFERENCE.md:
   both describe the rotation/attitude convention but are silent on the frame
   of the origin arguments specifically — not wrong, just silent — so neither
   needed a change.

Blast radius

git status --short after the change:

 M core/coords/transforms.py
 M tests/core/coords/test_transforms.py

Pure additions (2 files changed, 80 insertions(+), 0 deletions(-)). No other
.py file, no data file, and no figure byte touched.

Test plan

• Reproduced the pre-fix round-trip error via a scratch probe against
  this worktree's core (PYTHONPATH=$(pwd), confirmed the resolved
  transforms.py path was inside the worktree, not the main checkout):
  7.4727 m full 3D error / 6.2210 m on the Up axis alone.
• python -m pytest tests/core/coords -q → 64 passed.
• Mutation check: temporarily replaced
  enu_origin_body = -c_body_enu @ body_origin_enu with
  enu_origin_body = body_origin_enu in the new
  test_round_trip_with_offset_origin — it failed
  (Max absolute difference among violations: 7.17...) — then reverted,
  and the suite is green again (64 passed).
• python -m pytest tests/test_repo_conventions.py -q → 479 passed
  (pyflakes + convention ratchets).

🤖 Generated with Claude Code